Research And Implementation Of Robot Navigation Method Based On Vision Perception

Autonomous mobile robots are widely used in industries such as industry,aerospace,national defense,warehousing logistics,intelligent inspection and other fields.Mobile robots need to perform rescue operations,patrols,planetary exporation,material handling and other tasks,so mobile robots need to be able to complete autonomous navigation in a variety of static and dynamic environments.Visual sensors can perceive abundant environmental information and low cost.Autonomous navigation of robots by visual perception control has become a hot research topic in recent years.We design two robot navigation systems:robot navigation system based on scene segmentation and end-to-end robot navigation system.Robot navigation system based on scene segmentation uses ROS open source algorithm to construct environment map,locate and plan global path,and uses scene segmentation model to enhance the ability of robot to perceive scene.The system uses scene segmentation results,visual location algorithm and environment model to construct grid map in polar coordinates,and uses A star algorithm to plan path in grid map.The main innovations of the system are as follows:Firstly,we propose a large field of view scene sensing method,which uses multiple surround cameras to compensate for the small visual range of single camera.We use Scene Segmentation Model based on SegNet to accurately and efficiently perceive the environment Manual collection and annotation of environmental datasets improve model performance.Secondly,we use parallel projection to realize the mapping relationship between pixels coordinates and plane coordinates.The system builds the polar coordinate grid map to make the planning path continuous and smooth.Thirdly,the system uses the advantage of the large field of view perception to explore the path to help the robot move toward the target.The end-to-end robot navigation system use single camera image to control the robot.The main innovations of the system are as follows:Firstly,we propose a robot navigation model based on convolutional neural network.The model use input image to predict the steering Angle and collision probability of the current environment.Second,we manual collection and annotation the environmental images.The datasets contain two parts,one is used to train steering Angle,the other is used to train collision probability.Third,we design the model can be applied to tow-cost robots with limited hardware resources.The navigation model runs in NVIDIA Jetson TX2 and can control robot driving in real time.